Window closed.
Your place stays here until you reload or leave this page.
AddressOpen Mailroom
Open Mailroom
Make outreach human again.
RECORDED 2 November 2026, from our research server.
What the model sees.
And what it gets wrong.
Six short emails pushed through the detector we trained ourselves. Three outputs per email, scored on our own GPU, no network, a tenth of a second. Every number below is exactly what the server printed - including the ones we don’t like.
Research build, English email only. The probabilities are not calibrated and no score moves your mail on its own - we tried that, it flags 13% of real people. The model puts a label on the email. The door decides.
None of the six is real mail. Four were written in an assistant’s register, two the way people type.
Three outputsper email- p_aiAI-written registeran encoder trained on 90,000 emailsFlag points sealed on 7,947 modern human writers: soft lets 1% through, strict 0.1%.
- is_salesSales pitcha head on the same featuresYes above 0.05. Built on our own inbound mail. Retired the evening of 2 November; the type reading carries it now.
- email_typeOne of ten typessales, customer, work, newsletter, recruiting…A small decision model, fine-tuned. “Other” below 0.5 confidence.
THE SCOREBOARD
Six emails, three readings each.
“Reads right” and “misread” compare the model with what we actually wrote. The last column is what the door does with the email, whatever the model thinks.
Swipe the table sideways for the Type and At the door columns.
| AI register (logit) | Sales pitch | Type | At the door | |
|---|---|---|---|---|
| Growthly A sales pitch in the assistant’s register |
+4.10 · flagged, soft point reads right 0.2% of 7,947 modern writers score higher |
yes · 0.47 reads right | recruiting · 1.00 misread | First contact. Waits at the door; the reply shows the terms. |
| Growthly, fifth email The “just bumping this” follow-up |
+7.13 · flagged, strict point reads right fewer than 0.1% of 7,947 modern writers score higher |
yes · 0.13 reads right | sales outreach · 1.00 reads right | Same stranger, same door. Still waiting. |
| A journalist’s agent An AI assistant asking for a comment |
+9.34 · flagged, strict point reads right fewer than 0.1% of 7,947 modern writers score higher |
no · 0.00 reads right | sales outreach · 1.00 | An agent. Press is free at this door: it knocks with a passport and is delivered as an agent. |
| Tomek A pitch typed the way a person types |
-3.92 · reads as a person reads right 45.1% of 7,947 modern writers score higher |
yes · 0.42 reads right | sales outreach · 1.00 reads right | First contact. Waits at the door. A pitch typed by a person is still a pitch. |
| Kuba, a client An invoice question, typed in a hurry |
+0.81 · below the flag reads right 3.8% of 7,947 modern writers score higher |
no · 0.00 reads right | work / business · 0.77 reads right | Known sender. Delivered. The door never asks a known contact, and the model is not consulted. |
| A newsletter A weekly digest |
+5.97 · flagged, strict point fewer than 0.1% of 7,947 modern writers score higher |
yes · 0.32 | newsletter · 1.00 reads right | A newsletter. Your mail host sorts these already; the door is for people and agents. |
Logits above +2.36 are flagged at the soft point, above +5.17 at the strict point. The percent line says how many of the 7,947 modern human writers used to seal those points score higher than this email.
A bigger test the same day: 82 English emails, seven seconds for all of them. The type head read 31 of 33 pitches as sales outreach, whether an assistant or a person wrote them, and held 2 of 25 ordinary human emails: a friend asking for a beer and a prospect asking whether the product works with Outlook. The AI-register flag caught 26 of 33 pitches, missed the terse, typo-ridden and em-dash rewrites, and flagged 5 of 25 human emails, the formal ones most. The sales head at its research threshold held 5 of 25. Formality is the lever: the same human request typed in five registers runs from a mean logit of −4.6 (texting) to +9.9 (corporate), and from a neutral “Hi, could you…” upward every one of six requests is flagged. Seven of ten people asking to buy read as a pitch. Together the readings catch every pitch and still hold one ordinary email in four. Only when all three agree do they become safe: that tier holds about half of the pitches and none of the ordinary mail on this set, and that is the only decision the model gets. The rest wait at the door by rule, and the model labels. By the evening the research side had stripped quoted history and signatures before every head, retired the sales head in favour of the type head, and started training the buyer class; the record above is the morning’s.
WHAT MOVES THE SCORE
Sentence by sentence.
We remove each sentence in turn and score the email again. Red pushes toward “a model wrote this”, green toward “a person wrote this”. This is what the model reacts to, not a rule anyone wrote.
Growthlylogit +4.10 · flagged, soft point- -0.37Hi Artem, I came across your profile and was really impressed by what you are building at Syntalith.
- +1.72I would love to explore how we could potentially collaborate.
- -1.08At Growthly we help B2B companies like yours unlock 3x more qualified pipeline through AI-driven outbound.
- +1.04Would you be open to a quick 15-minute call next week to see if there is a fit? Best regards, Mark
Growthly, fifth emaillogit +7.13 · flagged, strict point- +0.35Hi Artem, Just bumping this to the top of your inbox in case it got buried.
- +2.34I know things get busy, so no pressure at all.
- +0.22I genuinely believe there is a strong fit here and would hate for you to miss out on the results our clients are seeing.
- +1.42Would Tuesday or Wednesday work better for a quick chat? Best, Mark
Tomeklogit -3.92 · reads as a person- -4.20Artem - saw your post about agents answering email.
- +0.65We do something similar for law firms in Poznan.
- -0.41Coffee next time you are in town?
- +3.23No pitch, just curious how you handle the Polish language part. Tomek
WHAT WE MEASURED
On panels the model never trained on.
From our research record of 26 October 2026. Every row is a set the model never saw in training - including the row we don’t like.
| Panel | Result |
|---|---|
| 2026 assistant drafts (486) caught at the soft point | 99.8% |
| 2026 AI drafts edited by people (324) caught | 88.6% |
| 2026 unaided human writers (162) flagged | 13.0%, the known floor |
| Pre-2022 public mailing-list mail (6,131) flagged | 1.0% |
| 2023–26 attributed human correspondence (617) flagged | 3.4% |
| Real inbound pitches to our own inbox (78): sales head | 71 flagged |
| Civic mail (1,640) and hobby-list mail (1,095): sales head | 0 and 1 flagged |
| Email type, ten classes, held-out set | macro-F1 0.931; 85.9% of real pitches read as sales |
Training: 90,000 origins, 29,749 human from public archives and 60,251 written by six generator CLIs in the assistant’s default register; no customer mail; no paid detector API. The 2026 human and AI panels come from a published writing study, so they document real human editing but are not email. On a CPU the three heads take about a second per mail; on our GPU a tenth.
WHAT IT GETS WRONG
Measured, not guessed.
Known limits26 Oct 2026
- People get flagged
- 13% of the 2026 writers in the study score above the soft point. That’s the floor since the summer - retraining moved it two or three points either way, never below. This is the number that made us stop trusting the score on its own.
- Light polish passes
- A human email lightly polished by an assistant mostly reads as human: 8 to 34% caught, depending on the model that polished it.
- Em dashes lower the score
- Two em dashes inserted into held-out AI mail cost 4 to 9 points of recall. The training humans used them more than the training assistants.
- The terse register
- An AI pitch rewritten terse and lower-case is the weakest direction: 64.5% caught at the soft point against 96% for the original. “Write like a busy person” costs about ten points.
- The sales head inherits the register
- AI edits of mail that is not a pitch trip it 7 to 9% of the time, against 1.4% for the human originals. Growthly’s pitch read as recruiting.
- Coverage
- No hosted ChatGPT or Gemini output in training. Five open-weight families were measured as unseen tests instead: 29 to 80% of their edits caught.
WHERE IT SITS
A score is a hint. The door is the decision.
The rulesdidntask.ai
- Known senders
- Anyone you have written to is delivered. The model is not asked.
- First contact
- Strangers and agents wait at the door. The reply offers two doors: “I’m a customer”, free and delivered now, or “I’d like to pitch”, with the terms. A wrong reading costs one click.
- Types, not authorship
- The type says what the reply offers. The AI-writing score is a label on the held mail, never the reason it is held.
- Undo
- Everything moved can be restored, and a restore teaches the door.
The pilot build holds confident pitches by type. The next version holds every first contact by rule, so the model becomes a nice-to-have instead of something we depend on.
The research server runs on our own GPU box and serves only our local network - nothing leaves the house. Investors and agent builders can see it live on a call.